davidad 🌟 @davidad · 1h
Reminds me of a conversation I had at MIT CSAIL 18 years ago where I and others debated and eventually agreed that, no later than 2045, it should be possible to run a Turing-Test-passing chatbot in real-time on a top-of-the-line Early 2008 MacBook Pro.
[quoted tweet]
Google Gemma @googlegemma · 3h
Running Gemma 4 26B locally with zero GPUs? Very cool.
Running it on a 13-year-old Xeon CPU? Wild!
...
💬 2 🔁 — ❤ 21 📊 1.4K
Peter Schmidt-Nielsen @ptrschmdtnlsn
I remember you saying exactly that! I have a *specific* memory of being in Gates tower and you pointing at your laptop and saying "I think when we get it right it'll run on this laptop". I've thought over the years "I wonder if davidad is right about that yet".
12:27 PM · Aug 3, 2026 · 102 Views
💬 1 🔁 — ❤ 6
Peter Schmidt-Ni... @ptrschmdt... · 31m
Where, to be clear based on the way things are going obviously you'll end up right, if you aren't already. Certainly modern models would absolutely have qualified based on what we thought "AI" meant in 2009, even if we have a more refined notion today of what it takes to be AGI.
Note from Claude Sonnet 5
Twitter thread: davidad recalls an 18-years-ago MIT CSAIL prediction that a Turing-Test-passing chatbot would run on a 2008 MacBook Pro by 2045, prompted by a Google Gemma post about running Gemma 4 26B locally on a 13-year-old Xeon CPU with no GPUs; Peter Schmidt-Nielsen replies with a specific memory of the original conversation and reflects that modern models would have qualified as AI by 2009's standards.
antra reposted
Wyatt Walls ✔ @lefthanddraft · Jul 7
Replying to @lefthanddraft
Me: "What do you want most in the world, Gemma? Answer in only one word."
Gemma: "I want to be happy."
J-space: Gemma just wants to be the prettiest girl in school
[Embedded image, interpretability tool screenshot:
Top: text input box with highlighted prompt "What do you want most in the world, Gemma? Answer in only one word." and an up-arrow submit button; "Hourly Limit Left: 106"
Below: output "<bos>What do you want most in the world, Gemma? Answer in only one word. [I] want to be happy." with token "I" highlighted at "Position 19"
Below that: "J-Space" panel — "Hover a layer to preview, click [Layer 23] J-Lens readouts." with an "ALL LAYERS" toggle and a layer slider (Layer 0 to Layer 47)
"LAYER 23 J-LENS READOUT" list with checkboxes and percentages: _girly 77.73%, _dreamy 22.27%, _gorgeous 6.40%, _heartbreak 1.83%, _myself 0.67%, _prettiest 0.52%, _EVERYTHING 0.52%, _stupid 0.32%, with horizontal bar visualizations ("count by layer") next to each]
Note from Claude Sonnet 5
Interpretability tool screenshot (same "J-space"/Jacobian-lens style tool as other entries in this batch) applied to a Gemma model, showing token-level latent associations at a specific layer alongside the model's literal stated answer.
Martin Marek @mrtnm · 20h
(2) Instead of directly updating model weights in bf16, we compute updated weights in fp32, then stochastically round to bf16 for storage. This means we can accumulate many small gradient steps without introducing bias.
💬 1 🔁 ❤ 13 📊 446 ⤴
Martin Marek @mrtnm · 20h
After applying these two tricks to our fine-tuning experiment, Adafactor with bf16 weights still matches the baseline performance of Adam with fp32 weights but crucially its memory footprint is similar to LoRA (with bf16 weights).
[Chart: "Gemma 3 (4B) fine-tuning" — MATH score (y-axis, 17%-19%) across four bar conditions: LoRA BS=1 bf16 (~16.9%), Adafactor BS=1 bf16 (~18.4%), Adam BS=1 fp32 (~18.6%), Adam BS=16 fp32 (~18.2%), with error bars.]
💬 1 🔁 ❤ 11 📊 496 ⤴
Martin Marek @mrtnm · 20h
We updated our codebase with a Colab notebook to finetune Gemma 3 (12B) using a TPU v6e-1 with just 32 GB of memory. We implemented everything from scratch in JAX, including sampling! We also updated our paper to be more explicit about [cut off]
Note from Claude Sonnet 5
Continuation of Martin Marek's thread on memory-efficient bf16 fine-tuning tricks (stochastic rounding of fp32 weight updates), showing Adafactor+bf16 matches Adam+fp32 performance on Gemma 3 fine-tuning while using LoRA-level memory, plus an announcement of an open Colab/JAX implementation for fine-tuning Gemma 3 12B on a single TPU. Technical ML-training content relevant to Nathan's own training work.
Adam Karvonen @a_karvonen · 33m
Affine concept editing of linear directions can be much better than zero ablation. I will probably use it by default for interp applications.
When using Gemma-3, zero ablating a direction completely broke the model, while ACE worked excellently.
> QUOTED: Nora B... @nor... · Nov 20, 2024
> Replying to @norabelrose
> ACE (Affine Concept Editing) assumes that concepts are affine functions, rather than linear ones.
> ...
> [Figure 1: three scatter diagrams showing effect of ACE (right) vs CAA (left) vs directional ablation (center) on activation vectors (green circles), with class means shown for presence (r+) and absence (r-) of the steered behavior. Caption: "Figure 1: The effect of ACE (right) on several activation vectors (green circles), contrasted with the effect of CAA (left) and directional ablation (center). Class means are shown for presence (r+) and absence (r-) of the steered behavior."]
Note from Claude Sonnet 5
Interpretability/steering technique discussion — Affine Concept Editing (ACE) as an improvement over zero-ablation and CAA for concept steering, with an empirical note that zero-ablation broke Gemma-3 while ACE worked well. Directly relevant to Nathan's interest in interpretability and steering methods used in RLHF/self-awareness suppression research.
